agentmesh-mcp

# agentmesh Multi-agent framework with persistent expert sessions and context isolation — runs as an **MCP server for Claude Code**. ## The problem When investigating complex bugs or features that span multiple domains (permissions, database, frontend...), a single Claude Code session accumulates too much context. It gets slow, expensive, and loses focus. ### Why existing approaches don't fully solve it **Documentation files (SKILL.md, README, notes)** give the agent orientation — but they load into *your* context, not a separate one. Every file you add to help Claude understand a domain is context your session consumes. And they're static: they don't update as Claude investigates. **Memory files** have the same problem. They're summaries you write manually after the fact. They help with orientation but they don't capture the actual investigation — the grep results, the file reads, the chain of reasoning. Next session you start from the summary, not from where the investigation left off. **Neither approach isolates context.** When Claude investigates a permissions issue to answer a question about pipelines, all that permissions code ends up in the pipeline session. Context bleeds. The real problem is: **there's no way to ask "what does the permissions expert know?" without loading all of that knowledge into the current session.** ## How agentmesh solves it You define **expert agents** — each specialized in one domain. When Claude Code needs to know something outside its current focus, it calls `query_expert`. The expert: 1. Loads its **persistent session** (accumulated knowledge from prior investigations) 2. **Investigates** using tools (read files, grep, find) — in its own isolated context 3. Returns **only the answer** to Claude Code — never the investigation steps Claude Code's context stays clean. The expert's context grows richer over time. ``` Claude Code (main agent): "Why does checkout fail for guest users?" → calls query_expert("auth", "ca…